Hello,
I've been having problems with Reiser 4 panicking for a few
months, and I've recently had time to investigate the matter. I've
created a program that can crash my system in a few minutes. It's
based on kmail's disk activity and consists of small, separated writes
to a file that is also mmapped.
=== scatteredwrites ===
#!/usr/bin/python
import os
import mmap
import optparse
parser = optparse.OptionParser(description=
"Creates a file in $CWD and performs a pattern of reads and writes to it in an "
"attempt to trigger fs bugs. The file is broken up into regions: for each "
"region the entire region is read, then some portion of it is written to."
"\nDistilled from kmail workload.")
parser.add_option("--region-size", dest="regionsize", default=65536,
type="int", help="Set region size to BYTES", metavar="BYTES")
parser.add_option("--region-count", dest="regioncount", default=2048,
type="int", help="Set number of regions to COUNT", metavar="COUNT")
parser.add_option("--write-offset", dest="writeoffset", default=0,
type="int", help="Offset write by BYTES in each region", metavar="BYTES")
parser.add_option("--write-size", dest="writesize", default=256,
type="int", help="Size of write in each region.", metavar="BYTES")
options, args = parser.parse_args()
f = open("scatteredwrites.%d.tmp" % (os.getpid()), "w+b")
try:
writestr = "A" * options.regionsize
for i in xrange(options.regioncount):
f.write(writestr)
f.close()
f = open("scatteredwrites.%d.tmp" % (os.getpid()), "r+b")
writestr = "B" * options.writesize
dummy = mmap.mmap(f.fileno(), options.regionsize * options.regioncount,
mmap.MAP_SHARED)
while True:
for i in xrange(options.regioncount):
f.seek(i * options.regionsize, 0)
f.read(options.regionsize)
f.seek(- options.regionsize + options.writeoffset,1)
f.write(writestr)
except KeyboardInterrupt:
os.unlink("scatteredwrites.%d.tmp" % (os.getpid()))
======
Without fs load this stress test rarely causes problems. But with five
instances running in parallel with five instances of a large grep (or
patch, or tar), my computer crashes on a timescale of 10 minutes.
I've also added a few patches to my kernel to help me debug the
problems I've been having:
diff -rupN a/fs/reiser4/page_cache.c b/fs/reiser4/page_cache.c
--- a/fs/reiser4/page_cache.c 2006-08-19 19:45:57.000000000 -0400
+++ b/fs/reiser4/page_cache.c 2006-08-19 20:23:43.000000000 -0400
@@ -489,12 +489,9 @@ static int can_hit_entd(reiser4_context
return 1;
if (ctx->super != s)
return 1;
- if (get_super_private(s)->entd.tsk == current)
- return 0;
- if (!lock_stack_isclean(&ctx->stack))
- return 0;
- if (ctx->trans->atom != NULL)
- return 0;
+ assert("ajw-1", get_super_private(s)->entd.tsk != current);
+ assert("ajw-2", lock_stack_isclean(&ctx->stack));
+ assert("ajw-3", ctx->trans->atom == NULL);
return 1;
}
diff -rupN 2.6.18-rc4-mm1/fs/reiser4/debug.c linux/fs/reiser4/debug.c
--- 2.6.18-rc4-mm1/fs/reiser4/debug.c 2006-08-18 19:21:13.000000000 -0400
+++ linux/fs/reiser4/debug.c 2006-08-18 19:24:35.000000000 -0400
@@ -56,6 +56,9 @@ static char panic_buf[REISER4_PANIC_MSG_
*/
static DEFINE_SPINLOCK(panic_guard);
+static void print_lock_counters(const char *prefix,
+ const reiser4_lock_counters_info * info);
+
/* Your best friend. Call it on each occasion. This is called by
fs/reiser4/debug.h:reiser4_panic(). */
void reiser4_do_panic(const char *format /* format string */ , ... /* rest */ )
@@ -74,6 +77,8 @@ void reiser4_do_panic(const char *format
vsnprintf(panic_buf, sizeof(panic_buf), format, args);
va_end(args);
printk(KERN_EMERG "reiser4 panicked cowardly: %s", panic_buf);
+ dump_stack();
+ print_lock_counters("",reiser4_lock_counters());
spin_unlock(&panic_guard);
/*
I've also added this bugfix by Alexander Zarochentsev <[EMAIL PROTECTED]>:
Index: linux-2.6-git/fs/reiser4/as_ops.c
===================================================================
--- linux-2.6-git.orig/fs/reiser4/as_ops.c
+++ linux-2.6-git/fs/reiser4/as_ops.c
@@ -350,6 +350,11 @@ int reiser4_releasepage(struct page *pag
if (PageDirty(page))
return 0;
+ /* extra page reference is used by reiser4 to protect
+ * jnode<->page link from this ->releasepage(). */
+ if (page_count(page) > 3)
+ return 0;
+
/* releasable() needs jnode lock, because it looks at the jnode fields
* and we need jload_lock here to avoid races with jload(). */
spin_lock_jnode(node);
Andrew Wade
